Zillow Offers

In April 2018, Zillow, best known for its Zestimate home valuation, announced that it would start buying and selling homes directly, beginning in Phoenix.¹ The idea, known as iBuying, was simple: use data and algorithms to make fast cash offers on houses, carry out light renovations, and resell them at a margin.

On 2 November 2021, Zillow announced that it would wind the business down. Its chief executive, Rich Barton, explained: “We’ve determined the unpredictability in forecasting home prices far exceeds what we anticipated and continuing to scale Zillow Offers would result in too much earnings and balance-sheet volatility.”²

Scale arrived before certainty

The numbers in Zillow’s third-quarter 2021 letter to shareholders show how quickly the business had grown. In that quarter alone, Zillow Offers bought 9,680 homes and sold 3,032. It ended the quarter with 9,790 homes in inventory and a further 8,172 under contract.³

Zillow recorded a $304 million inventory write-down in the quarter, because it had bought homes “at higher prices than our current estimates of future selling prices”, and said it expected further losses of $240 million to $265 million in the fourth quarter. The Homes segment lost $422 million before tax in the third quarter.^{3 2} About a quarter of the company’s workforce was cut.²

Why it failed

Zillow’s own explanation was blunt: “We have been unable to accurately forecast future home prices at different times in both directions.” It also pointed to “industry-wide renovation labor supply shortages”, which slowed the work needed to resell homes and pushed sales into later quarters.³

Amit Seru, professor of finance at Stanford Graduate School of Business, has offered a more structural explanation.⁴ Three points stand out.

The model could not see what sellers could. “There are a lot of features and attributes of a home that the model doesn’t capture,” Seru notes. Owners know about the damp patch, the noisy road or the awkward layout. An algorithm making fast offers on averages invites what economists call adverse selection: the people most keen to accept are often those whose homes are worth less than the model thinks.

Errors only cancel out on uniform portfolios. Seru argues that pricing errors tend to wash out across large portfolios of similar, “cookie-cutter” homes, but not once a buyer moves into more varied properties, as Zillow did.

Growth pressure pushed bids beyond the model. “To compete, Zillow started bidding more for cookie-cutter homes than what their algorithmic model predicted for such homes,” Seru says. As a late entrant chasing market share, Zillow was taking on more risk at the moment its market was becoming harder to predict.

What happened next

Zillow sold its remaining inventory over the following quarters and refocused on its core advertising and agent business. Other iBuyers continued to operate, which reinforces the point: the failure was not that algorithmic home buying is impossible, but that this business scaled its exposure faster than it could learn whether its predictions held.

Lessons for leaders

1. Scale exposure in step with evidence. Zillow was buying nearly 10,000 homes a quarter on the strength of a model whose errors only became clear months later, when the houses were resold. The cost of a wrong prediction is set by how much you have committed before you find out.
2. Watch for commercial pressure overriding the model. When targets push teams to bid above what the model recommends, the organisation is no longer using AI to decide; it is using AI to justify. Someone must be accountable for spotting that drift.
3. Ask what the model cannot see, and who can. In any market where the other party knows more than your data does, assume your fastest-accepted decisions are your worst ones, and design checks accordingly.

Questions to ask in your organisation

1. How long is the gap between our AI making a decision and us finding out whether it was right, and how much do we commit in that time?
2. Where have targets or competitive pressure led us to adjust, override or loosen a model’s recommendations, and who signed that off?
3. Which of our automated decisions are most exposed to people who know more than our data, such as customers, sellers or applicants?